WoodScape: A multi-task, multi-camera fisheye dataset for autonomous driving- Supplementary Material

Senthil Yogamani, Ciaran Hughes, Jonathan Horgan, Ganesh Sistu, Padraig Varley, Derek O’Dea,Michal Uricar, Stefan Milz, Martin Simon, Karl Amende, Christian Witt, Hazem Rashed,

Sumanth Chennupati, Sanjaya Nayak, Saquib Mansoor, Xavier Perroton, Patrick Perezhttps://github.com/valeoai/WoodScape firstname.lastname@valeo.com

1. Accuracy breakdown of semantic tasksTable 1 summarizes the statistics of semantic segmenta-

tion and bounding box detection tasks in our 10k dataset.The samples are carefully chosen to ensure a good distri-bution of object. We refined the instance level annotationsand generated bounding box annotations in PASCAL VOCstyle. It is be noted that these are preliminary results and fi-nal results may vary due to sampling variations and trainingconfiguration. When the baseline network is released in ourgithub, corresponding accuracies will be provided too.

Table 1: Summary of results of baseline experiments. Fromtop to bottom: (1) Distribution of frames across differentcamera positions, (2) Number of 2D box samples for eachclass and its corresponding mAP score using FasterRCNNand (3) Class-wise accuracies of semantic segmentation net-work.

Camera DistributionsFront Rear Left Right3500 3500 750 750

Bounding Box Object Distributions4-Wheeler 2-Wheeler Person Animals

56.5k 8.7k 54k 0.5kFasterRCNN mAP@0.5IOU

0.63 0.17 0.44 0.01k

Semantic Segmentation (ENet) IOU scores4-Wheeler 2-Wheeler Person Road Void

0.55 0.22 0.43 0.95 0.42

2. Annotated class hierarchyHere we provide the complete class hierarchy for the an-

notation of the WoodScape dataset:

• sidewalk

– curb

– free space

– sidewalk

• road

– road

– road surface

• four wheelers

– dynamic car

– car

– dynamic van

– van

– caravan

• four wheelers heavy

– dynamic bus

– dynamic truck

– dynamic rv

– dynamic trailer

– dynamic on rails

– bus

– truck

– trailer

– train/tram

• four wheelers groups

– dynamic vehicle group

– grouped vehicles

• two wheelers

– dynamic bicycle

1

– dynamic motorcycle

– bicycle

– motorcycle

• two wheelers groups

– dynamic cycle group

• person

– dynamic person

– person

• rider

– dynamic rider

– rider

• person group

– dynamic person group

• lane marks

– road lane line

– road parking line

– lane marking

– parking marking

– zebra crossing

– parking line

• ground markings

– road reserved parking

– empty reserved

– other ground marking

– zebra crossing

• catseye bottsdots

– road botts dots

– road group botts dots

• animals

– dynamic animal

– dynamic animal group

– animal

– grouped animals

• structure

– fence guardrail

– structure

– construction

– fence

• object

– dynamic object

– pole-post

– pole

• traffic signs

– other traffic sign

– other traffic sign indistinguishable

– red traffic light

– yellow traffic light

– green traffic light

– unknown traffic light

– unknown traffic light

– traffic light red

– traffic sign indistingushable

– traffic light green

– traffic sign

– traffic light yellow

• nature

– terrain

– nature

• sky

– sky

• void

– other nosight

– other blur

– ego vehicle

– green strip

– dynamic

– background

– void

– cats eyes and botts dots

– movable object

– grouped botts dots

– other wheeled transport

– grouped pedestrian and animals

3. Synthetic DataThe synthetic data is generated using a commercial phys-

ically accurate simulation engine ANYVERSE. Our cameramodel including intrinsics and extrinsics were incorporatedas illustrated in Figure 1. Annotation for instance segmen-tation and depth estimation will be provided.

Figure 1: Illustration of synthetic images generated using our fisheye camera model.